orient-cli 0.1.0 → 0.2.0

package/node_modules/@orient-cli/ai/dist/providers/google-gemini-cli.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"google-gemini-cli.d.ts","sourceRoot":"","sources":["../../src/providers/google-gemini-cli.ts"],"names":[],"mappings":"AAAA;;;;GAIG;AAEH,OAAO,KAAK,EAAE,OAAO,EAAE,cAAc,EAAE,MAAM,eAAe,CAAC;AAE7D,OAAO,KAAK,EAGX,OAAO,EACP,KAAK,EACL,mBAAmB,EACnB,cAAc,EACd,aAAa,EAMb,MAAM,aAAa,CAAC;AAGrB,OAAO,EAEN,YAAY,EAGZ,aAAa,EAEb,MAAM,oBAAoB,CAAC;AAG5B;;;GAGG;AACH,MAAM,MAAM,mBAAmB,GAAG,4BAA4B,GAAG,SAAS,GAAG,KAAK,GAAG,QAAQ,GAAG,MAAM,CAAC;AAEvG,MAAM,WAAW,sBAAuB,SAAQ,aAAa;IAC5D,UAAU,CAAC,EAAE,MAAM,GAAG,MAAM,GAAG,KAAK,CAAC;IACrC;;;;;;OAMG;IACH,QAAQ,CAAC,EAAE;QACV,OAAO,EAAE,OAAO,CAAC;QACjB,4DAA4D;QAC5D,YAAY,CAAC,EAAE,MAAM,CAAC;QACtB,qGAAqG;QACrG,KAAK,CAAC,EAAE,mBAAmB,CAAC;KAC5B,CAAC;IACF,SAAS,CAAC,EAAE,MAAM,CAAC;CACnB;AAgDD;;;;;;;;GAQG;AACH,wBAAgB,iBAAiB,CAAC,SAAS,EAAE,MAAM,EAAE,QAAQ,CAAC,EAAE,QAAQ,GAAG,OAAO,GAAG,MAAM,GAAG,SAAS,CAyFtG;AA6DD,UAAU,sBAAsB;IAC/B,OAAO,EAAE,MAAM,CAAC;IAChB,KAAK,EAAE,MAAM,CAAC;IACd,OAAO,EAAE;QACR,QAAQ,EAAE,OAAO,EAAE,CAAC;QACpB,SAAS,CAAC,EAAE,MAAM,CAAC;QACnB,iBAAiB,CAAC,EAAE;YAAE,IAAI,CAAC,EAAE,MAAM,CAAC;YAAC,KAAK,EAAE;gBAAE,IAAI,EAAE,MAAM,CAAA;aAAE,EAAE,CAAA;SAAE,CAAC;QACjE,gBAAgB,CAAC,EAAE;YAClB,eAAe,CAAC,EAAE,MAAM,CAAC;YACzB,WAAW,CAAC,EAAE,MAAM,CAAC;YACrB,cAAc,CAAC,EAAE,cAAc,CAAC;SAChC,CAAC;QACF,KAAK,CAAC,EAAE,UAAU,CAAC,OAAO,YAAY,CAAC,CAAC;QACxC,UAAU,CAAC,EAAE;YACZ,qBAAqB,EAAE;gBACtB,IAAI,EAAE,UAAU,CAAC,OAAO,aAAa,CAAC,CAAC;aACvC,CAAC;SACF,CAAC;KACF,CAAC;IACF,WAAW,CAAC,EAAE,MAAM,CAAC;IACrB,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB,SAAS,CAAC,EAAE,MAAM,CAAC;CACnB;AAiCD,eAAO,MAAM,qBAAqB,EAAE,cAAc,CAAC,mBAAmB,EAAE,sBAAsB,CAse7F,CAAC;AAEF,eAAO,MAAM,2BAA2B,EAAE,cAAc,CAAC,mBAAmB,EAAE,mBAAmB,CAqDhG,CAAC;AAEF,wBAAgB,YAAY,CAC3B,KAAK,EAAE,KAAK,CAAC,mBAAmB,CAAC,EACjC,OAAO,EAAE,OAAO,EAChB,SAAS,EAAE,MAAM,EACjB,OAAO,GAAE,sBAA2B,EACpC,aAAa,UAAQ,GACnB,sBAAsB,CA8ExB","sourcesContent":["/**\n * Google Gemini CLI / Antigravity provider.\n * Shared implementation for both google-gemini-cli and google-antigravity providers.\n * Uses the Cloud Code Assist API endpoint to access Gemini and Claude models.\n */\n\nimport type { Content, ThinkingConfig } from \"@google/genai\";\nimport { calculateCost } from \"../models.js\";\nimport type {\n\tApi,\n\tAssistantMessage,\n\tContext,\n\tModel,\n\tSimpleStreamOptions,\n\tStreamFunction,\n\tStreamOptions,\n\tTextContent,\n\tThinkingBudgets,\n\tThinkingContent,\n\tThinkingLevel,\n\tToolCall,\n} from \"../types.js\";\nimport { AssistantMessageEventStream } from \"../utils/event-stream.js\";\nimport { sanitizeSurrogates } from \"../utils/sanitize-unicode.js\";\nimport {\n\tconvertMessages,\n\tconvertTools,\n\tisThinkingPart,\n\tmapStopReasonString,\n\tmapToolChoice,\n\tretainThoughtSignature,\n} from \"./google-shared.js\";\nimport { buildBaseOptions, clampReasoning } from \"./simple-options.js\";\n\n/**\n * Thinking level for Gemini 3 models.\n * Mirrors Google's ThinkingLevel enum values.\n */\nexport type GoogleThinkingLevel = \"THINKING_LEVEL_UNSPECIFIED\" | \"MINIMAL\" | \"LOW\" | \"MEDIUM\" | \"HIGH\";\n\nexport interface GoogleGeminiCliOptions extends StreamOptions {\n\ttoolChoice?: \"auto\" | \"none\" | \"any\";\n\t/**\n\t * Thinking/reasoning configuration.\n\t * - Gemini 2.x models: use `budgetTokens` to set the thinking budget\n\t * - Gemini 3 models (gemini-3-pro-*, gemini-3-flash-*): use `level` instead\n\t *\n\t * When using `streamSimple`, this is handled automatically based on the model.\n\t */\n\tthinking?: {\n\t\tenabled: boolean;\n\t\t/** Thinking budget in tokens. Use for Gemini 2.x models. */\n\t\tbudgetTokens?: number;\n\t\t/** Thinking level. Use for Gemini 3 models (LOW/HIGH for Pro, MINIMAL/LOW/MEDIUM/HIGH for Flash). */\n\t\tlevel?: GoogleThinkingLevel;\n\t};\n\tprojectId?: string;\n}\n\nconst DEFAULT_ENDPOINT = \"https://cloudcode-pa.googleapis.com\";\nconst ANTIGRAVITY_DAILY_ENDPOINT = \"https://daily-cloudcode-pa.sandbox.googleapis.com\";\nconst ANTIGRAVITY_AUTOPUSH_ENDPOINT = \"https://autopush-cloudcode-pa.sandbox.googleapis.com\";\nconst ANTIGRAVITY_ENDPOINT_FALLBACKS = [\n\tANTIGRAVITY_DAILY_ENDPOINT,\n\tANTIGRAVITY_AUTOPUSH_ENDPOINT,\n\tDEFAULT_ENDPOINT,\n] as const;\n// Headers for Gemini CLI (prod endpoint)\nconst GEMINI_CLI_HEADERS = {\n\t\"User-Agent\": \"google-cloud-sdk vscode_cloudshelleditor/0.1\",\n\t\"X-Goog-Api-Client\": \"gl-node/22.17.0\",\n\t\"Client-Metadata\": JSON.stringify({\n\t\tideType: \"IDE_UNSPECIFIED\",\n\t\tplatform: \"PLATFORM_UNSPECIFIED\",\n\t\tpluginType: \"GEMINI\",\n\t}),\n};\n\n// Headers for Antigravity (sandbox endpoint) - requires specific User-Agent\nconst DEFAULT_ANTIGRAVITY_VERSION = \"1.21.9\";\n\nfunction getAntigravityHeaders() {\n\tconst version = process.env.PI_AI_ANTIGRAVITY_VERSION || DEFAULT_ANTIGRAVITY_VERSION;\n\treturn {\n\t\t\"User-Agent\": `antigravity/${version} darwin/arm64`,\n\t};\n}\n\n// Antigravity system instruction (compact version from CLIProxyAPI).\nconst ANTIGRAVITY_SYSTEM_INSTRUCTION =\n\t\"You are Antigravity, a powerful agentic AI coding assistant designed by the Google Deepmind team working on Advanced Agentic Coding.\" +\n\t\"You are pair programming with a USER to solve their coding task. The task may require creating a new codebase, modifying or debugging an existing codebase, or simply answering a question.\" +\n\t\"**Absolute paths only**\" +\n\t\"**Proactiveness**\";\n\n// Counter for generating unique tool call IDs\nlet toolCallCounter = 0;\n\n// Retry configuration\nconst MAX_RETRIES = 3;\nconst BASE_DELAY_MS = 1000;\nconst MAX_EMPTY_STREAM_RETRIES = 2;\nconst EMPTY_STREAM_BASE_DELAY_MS = 500;\nconst CLAUDE_THINKING_BETA_HEADER = \"interleaved-thinking-2025-05-14\";\n\n/**\n * Extract retry delay from Gemini error response (in milliseconds).\n * Checks headers first (Retry-After, x-ratelimit-reset, x-ratelimit-reset-after),\n * then parses body patterns like:\n * - \"Your quota will reset after 39s\"\n * - \"Your quota will reset after 18h31m10s\"\n * - \"Please retry in Xs\" or \"Please retry in Xms\"\n * - \"retryDelay\": \"34.074824224s\" (JSON field)\n */\nexport function extractRetryDelay(errorText: string, response?: Response | Headers): number | undefined {\n\tconst normalizeDelay = (ms: number): number | undefined => (ms > 0 ? Math.ceil(ms + 1000) : undefined);\n\n\tconst headers = response instanceof Headers ? response : response?.headers;\n\tif (headers) {\n\t\tconst retryAfter = headers.get(\"retry-after\");\n\t\tif (retryAfter) {\n\t\t\tconst retryAfterSeconds = Number(retryAfter);\n\t\t\tif (Number.isFinite(retryAfterSeconds)) {\n\t\t\t\tconst delay = normalizeDelay(retryAfterSeconds * 1000);\n\t\t\t\tif (delay !== undefined) {\n\t\t\t\t\treturn delay;\n\t\t\t\t}\n\t\t\t}\n\t\t\tconst retryAfterDate = new Date(retryAfter);\n\t\t\tconst retryAfterMs = retryAfterDate.getTime();\n\t\t\tif (!Number.isNaN(retryAfterMs)) {\n\t\t\t\tconst delay = normalizeDelay(retryAfterMs - Date.now());\n\t\t\t\tif (delay !== undefined) {\n\t\t\t\t\treturn delay;\n\t\t\t\t}\n\t\t\t}\n\t\t}\n\n\t\tconst rateLimitReset = headers.get(\"x-ratelimit-reset\");\n\t\tif (rateLimitReset) {\n\t\t\tconst resetSeconds = Number.parseInt(rateLimitReset, 10);\n\t\t\tif (!Number.isNaN(resetSeconds)) {\n\t\t\t\tconst delay = normalizeDelay(resetSeconds * 1000 - Date.now());\n\t\t\t\tif (delay !== undefined) {\n\t\t\t\t\treturn delay;\n\t\t\t\t}\n\t\t\t}\n\t\t}\n\n\t\tconst rateLimitResetAfter = headers.get(\"x-ratelimit-reset-after\");\n\t\tif (rateLimitResetAfter) {\n\t\t\tconst resetAfterSeconds = Number(rateLimitResetAfter);\n\t\t\tif (Number.isFinite(resetAfterSeconds)) {\n\t\t\t\tconst delay = normalizeDelay(resetAfterSeconds * 1000);\n\t\t\t\tif (delay !== undefined) {\n\t\t\t\t\treturn delay;\n\t\t\t\t}\n\t\t\t}\n\t\t}\n\t}\n\n\t// Pattern 1: \"Your quota will reset after ...\" (formats: \"18h31m10s\", \"10m15s\", \"6s\", \"39s\")\n\tconst durationMatch = errorText.match(/reset after (?:(\\d+)h)?(?:(\\d+)m)?(\\d+(?:\\.\\d+)?)s/i);\n\tif (durationMatch) {\n\t\tconst hours = durationMatch[1] ? parseInt(durationMatch[1], 10) : 0;\n\t\tconst minutes = durationMatch[2] ? parseInt(durationMatch[2], 10) : 0;\n\t\tconst seconds = parseFloat(durationMatch[3]);\n\t\tif (!Number.isNaN(seconds)) {\n\t\t\tconst totalMs = ((hours * 60 + minutes) * 60 + seconds) * 1000;\n\t\t\tconst delay = normalizeDelay(totalMs);\n\t\t\tif (delay !== undefined) {\n\t\t\t\treturn delay;\n\t\t\t}\n\t\t}\n\t}\n\n\t// Pattern 2: \"Please retry in X[ms|s]\"\n\tconst retryInMatch = errorText.match(/Please retry in ([0-9.]+)(ms|s)/i);\n\tif (retryInMatch?.[1]) {\n\t\tconst value = parseFloat(retryInMatch[1]);\n\t\tif (!Number.isNaN(value) && value > 0) {\n\t\t\tconst ms = retryInMatch[2].toLowerCase() === \"ms\" ? value : value * 1000;\n\t\t\tconst delay = normalizeDelay(ms);\n\t\t\tif (delay !== undefined) {\n\t\t\t\treturn delay;\n\t\t\t}\n\t\t}\n\t}\n\n\t// Pattern 3: \"retryDelay\": \"34.074824224s\" (JSON field in error details)\n\tconst retryDelayMatch = errorText.match(/\"retryDelay\":\\s*\"([0-9.]+)(ms|s)\"/i);\n\tif (retryDelayMatch?.[1]) {\n\t\tconst value = parseFloat(retryDelayMatch[1]);\n\t\tif (!Number.isNaN(value) && value > 0) {\n\t\t\tconst ms = retryDelayMatch[2].toLowerCase() === \"ms\" ? value : value * 1000;\n\t\t\tconst delay = normalizeDelay(ms);\n\t\t\tif (delay !== undefined) {\n\t\t\t\treturn delay;\n\t\t\t}\n\t\t}\n\t}\n\n\treturn undefined;\n}\n\nfunction needsClaudeThinkingBetaHeader(model: Model<\"google-gemini-cli\">): boolean {\n\treturn model.provider === \"google-antigravity\" && model.id.startsWith(\"claude-\") && model.reasoning;\n}\n\nfunction isGemini3ProModel(modelId: string): boolean {\n\treturn /gemini-3(?:\\.1)?-pro/.test(modelId.toLowerCase());\n}\n\nfunction isGemini3FlashModel(modelId: string): boolean {\n\treturn /gemini-3(?:\\.1)?-flash/.test(modelId.toLowerCase());\n}\n\nfunction isGemini3Model(modelId: string): boolean {\n\treturn isGemini3ProModel(modelId) || isGemini3FlashModel(modelId);\n}\n\n/**\n * Check if an error is retryable (rate limit, server error, network error, etc.)\n */\nfunction isRetryableError(status: number, errorText: string): boolean {\n\tif (status === 429 || status === 500 || status === 502 || status === 503 || status === 504) {\n\t\treturn true;\n\t}\n\treturn /resource.?exhausted|rate.?limit|overloaded|service.?unavailable|other.?side.?closed/i.test(errorText);\n}\n\n/**\n * Extract a clean, user-friendly error message from Google API error response.\n * Parses JSON error responses and returns just the message field.\n */\nfunction extractErrorMessage(errorText: string): string {\n\ttry {\n\t\tconst parsed = JSON.parse(errorText) as { error?: { message?: string } };\n\t\tif (parsed.error?.message) {\n\t\t\treturn parsed.error.message;\n\t\t}\n\t} catch {\n\t\t// Not JSON, return as-is\n\t}\n\treturn errorText;\n}\n\n/**\n * Sleep for a given number of milliseconds, respecting abort signal.\n */\nfunction sleep(ms: number, signal?: AbortSignal): Promise<void> {\n\treturn new Promise((resolve, reject) => {\n\t\tif (signal?.aborted) {\n\t\t\treject(new Error(\"Request was aborted\"));\n\t\t\treturn;\n\t\t}\n\t\tconst timeout = setTimeout(resolve, ms);\n\t\tsignal?.addEventListener(\"abort\", () => {\n\t\t\tclearTimeout(timeout);\n\t\t\treject(new Error(\"Request was aborted\"));\n\t\t});\n\t});\n}\n\ninterface CloudCodeAssistRequest {\n\tproject: string;\n\tmodel: string;\n\trequest: {\n\t\tcontents: Content[];\n\t\tsessionId?: string;\n\t\tsystemInstruction?: { role?: string; parts: { text: string }[] };\n\t\tgenerationConfig?: {\n\t\t\tmaxOutputTokens?: number;\n\t\t\ttemperature?: number;\n\t\t\tthinkingConfig?: ThinkingConfig;\n\t\t};\n\t\ttools?: ReturnType<typeof convertTools>;\n\t\ttoolConfig?: {\n\t\t\tfunctionCallingConfig: {\n\t\t\t\tmode: ReturnType<typeof mapToolChoice>;\n\t\t\t};\n\t\t};\n\t};\n\trequestType?: string;\n\tuserAgent?: string;\n\trequestId?: string;\n}\n\ninterface CloudCodeAssistResponseChunk {\n\tresponse?: {\n\t\tcandidates?: Array<{\n\t\t\tcontent?: {\n\t\t\t\trole: string;\n\t\t\t\tparts?: Array<{\n\t\t\t\t\ttext?: string;\n\t\t\t\t\tthought?: boolean;\n\t\t\t\t\tthoughtSignature?: string;\n\t\t\t\t\tfunctionCall?: {\n\t\t\t\t\t\tname: string;\n\t\t\t\t\t\targs: Record<string, unknown>;\n\t\t\t\t\t\tid?: string;\n\t\t\t\t\t};\n\t\t\t\t}>;\n\t\t\t};\n\t\t\tfinishReason?: string;\n\t\t}>;\n\t\tusageMetadata?: {\n\t\t\tpromptTokenCount?: number;\n\t\t\tcandidatesTokenCount?: number;\n\t\t\tthoughtsTokenCount?: number;\n\t\t\ttotalTokenCount?: number;\n\t\t\tcachedContentTokenCount?: number;\n\t\t};\n\t\tmodelVersion?: string;\n\t\tresponseId?: string;\n\t};\n\ttraceId?: string;\n}\n\nexport const streamGoogleGeminiCli: StreamFunction<\"google-gemini-cli\", GoogleGeminiCliOptions> = (\n\tmodel: Model<\"google-gemini-cli\">,\n\tcontext: Context,\n\toptions?: GoogleGeminiCliOptions,\n): AssistantMessageEventStream => {\n\tconst stream = new AssistantMessageEventStream();\n\n\t(async () => {\n\t\tconst output: AssistantMessage = {\n\t\t\trole: \"assistant\",\n\t\t\tcontent: [],\n\t\t\tapi: \"google-gemini-cli\" as Api,\n\t\t\tprovider: model.provider,\n\t\t\tmodel: model.id,\n\t\t\tusage: {\n\t\t\t\tinput: 0,\n\t\t\t\toutput: 0,\n\t\t\t\tcacheRead: 0,\n\t\t\t\tcacheWrite: 0,\n\t\t\t\ttotalTokens: 0,\n\t\t\t\tcost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },\n\t\t\t},\n\t\t\tstopReason: \"stop\",\n\t\t\ttimestamp: Date.now(),\n\t\t};\n\n\t\ttry {\n\t\t\t// apiKey is JSON-encoded: { token, projectId }\n\t\t\tconst apiKeyRaw = options?.apiKey;\n\t\t\tif (!apiKeyRaw) {\n\t\t\t\tthrow new Error(\"Google Cloud Code Assist requires OAuth authentication. Use /login to authenticate.\");\n\t\t\t}\n\n\t\t\tlet accessToken: string;\n\t\t\tlet projectId: string;\n\n\t\t\ttry {\n\t\t\t\tconst parsed = JSON.parse(apiKeyRaw) as { token: string; projectId: string };\n\t\t\t\taccessToken = parsed.token;\n\t\t\t\tprojectId = parsed.projectId;\n\t\t\t} catch {\n\t\t\t\tthrow new Error(\"Invalid Google Cloud Code Assist credentials. Use /login to re-authenticate.\");\n\t\t\t}\n\n\t\t\tif (!accessToken || !projectId) {\n\t\t\t\tthrow new Error(\"Missing token or projectId in Google Cloud credentials. Use /login to re-authenticate.\");\n\t\t\t}\n\n\t\t\tconst isAntigravity = model.provider === \"google-antigravity\";\n\t\t\tconst baseUrl = model.baseUrl?.trim();\n\t\t\tconst endpoints = baseUrl ? [baseUrl] : isAntigravity ? ANTIGRAVITY_ENDPOINT_FALLBACKS : [DEFAULT_ENDPOINT];\n\n\t\t\tlet requestBody = buildRequest(model, context, projectId, options, isAntigravity);\n\t\t\tconst nextRequestBody = await options?.onPayload?.(requestBody, model);\n\t\t\tif (nextRequestBody !== undefined) {\n\t\t\t\trequestBody = nextRequestBody as CloudCodeAssistRequest;\n\t\t\t}\n\t\t\tconst headers = isAntigravity ? getAntigravityHeaders() : GEMINI_CLI_HEADERS;\n\n\t\t\tconst requestHeaders = {\n\t\t\t\tAuthorization: `Bearer ${accessToken}`,\n\t\t\t\t\"Content-Type\": \"application/json\",\n\t\t\t\tAccept: \"text/event-stream\",\n\t\t\t\t...headers,\n\t\t\t\t...(needsClaudeThinkingBetaHeader(model) ? { \"anthropic-beta\": CLAUDE_THINKING_BETA_HEADER } : {}),\n\t\t\t\t...options?.headers,\n\t\t\t};\n\t\t\tconst requestBodyJson = JSON.stringify(requestBody);\n\n\t\t\t// Fetch with retry logic for rate limits, transient errors, and endpoint fallbacks.\n\t\t\t// On 403/404, immediately try the next endpoint (no delay).\n\t\t\t// On 429/5xx, retry with backoff on the same or next endpoint.\n\t\t\tlet response: Response | undefined;\n\t\t\tlet lastError: Error | undefined;\n\t\t\tlet requestUrl: string | undefined;\n\t\t\tlet endpointIndex = 0;\n\n\t\t\tfor (let attempt = 0; attempt <= MAX_RETRIES; attempt++) {\n\t\t\t\tif (options?.signal?.aborted) {\n\t\t\t\t\tthrow new Error(\"Request was aborted\");\n\t\t\t\t}\n\n\t\t\t\ttry {\n\t\t\t\t\tconst endpoint = endpoints[endpointIndex];\n\t\t\t\t\trequestUrl = `${endpoint}/v1internal:streamGenerateContent?alt=sse`;\n\t\t\t\t\tresponse = await fetch(requestUrl, {\n\t\t\t\t\t\tmethod: \"POST\",\n\t\t\t\t\t\theaders: requestHeaders,\n\t\t\t\t\t\tbody: requestBodyJson,\n\t\t\t\t\t\tsignal: options?.signal,\n\t\t\t\t\t});\n\n\t\t\t\t\tif (response.ok) {\n\t\t\t\t\t\tbreak; // Success, exit retry loop\n\t\t\t\t\t}\n\n\t\t\t\t\tconst errorText = await response.text();\n\n\t\t\t\t\t// On 403/404, cascade to the next endpoint immediately (no delay)\n\t\t\t\t\tif ((response.status === 403 || response.status === 404) && endpointIndex < endpoints.length - 1) {\n\t\t\t\t\t\tendpointIndex++;\n\t\t\t\t\t\tcontinue;\n\t\t\t\t\t}\n\n\t\t\t\t\t// Check if retryable (429, 5xx, network patterns)\n\t\t\t\t\tif (attempt < MAX_RETRIES && isRetryableError(response.status, errorText)) {\n\t\t\t\t\t\t// Advance endpoint if possible\n\t\t\t\t\t\tif (endpointIndex < endpoints.length - 1) {\n\t\t\t\t\t\t\tendpointIndex++;\n\t\t\t\t\t\t}\n\n\t\t\t\t\t\t// Use server-provided delay or exponential backoff\n\t\t\t\t\t\tconst serverDelay = extractRetryDelay(errorText, response);\n\t\t\t\t\t\tconst delayMs = serverDelay ?? BASE_DELAY_MS * 2 ** attempt;\n\n\t\t\t\t\t\t// Check if server delay exceeds max allowed (default: 60s)\n\t\t\t\t\t\tconst maxDelayMs = options?.maxRetryDelayMs ?? 60000;\n\t\t\t\t\t\tif (maxDelayMs > 0 && serverDelay && serverDelay > maxDelayMs) {\n\t\t\t\t\t\t\tconst delaySeconds = Math.ceil(serverDelay / 1000);\n\t\t\t\t\t\t\tthrow new Error(\n\t\t\t\t\t\t\t\t`Server requested ${delaySeconds}s retry delay (max: ${Math.ceil(maxDelayMs / 1000)}s). ${extractErrorMessage(errorText)}`,\n\t\t\t\t\t\t\t);\n\t\t\t\t\t\t}\n\n\t\t\t\t\t\tawait sleep(delayMs, options?.signal);\n\t\t\t\t\t\tcontinue;\n\t\t\t\t\t}\n\n\t\t\t\t\t// Not retryable or max retries exceeded\n\t\t\t\t\tthrow new Error(`Cloud Code Assist API error (${response.status}): ${extractErrorMessage(errorText)}`);\n\t\t\t\t} catch (error) {\n\t\t\t\t\t// Check for abort - fetch throws AbortError, our code throws \"Request was aborted\"\n\t\t\t\t\tif (error instanceof Error) {\n\t\t\t\t\t\tif (error.name === \"AbortError\" || error.message === \"Request was aborted\") {\n\t\t\t\t\t\t\tthrow new Error(\"Request was aborted\");\n\t\t\t\t\t\t}\n\t\t\t\t\t}\n\t\t\t\t\t// Extract detailed error message from fetch errors (Node includes cause)\n\t\t\t\t\tlastError = error instanceof Error ? error : new Error(String(error));\n\t\t\t\t\tif (lastError.message === \"fetch failed\" && lastError.cause instanceof Error) {\n\t\t\t\t\t\tlastError = new Error(`Network error: ${lastError.cause.message}`);\n\t\t\t\t\t}\n\t\t\t\t\t// Network errors are retryable\n\t\t\t\t\tif (attempt < MAX_RETRIES) {\n\t\t\t\t\t\tconst delayMs = BASE_DELAY_MS * 2 ** attempt;\n\t\t\t\t\t\tawait sleep(delayMs, options?.signal);\n\t\t\t\t\t\tcontinue;\n\t\t\t\t\t}\n\t\t\t\t\tthrow lastError;\n\t\t\t\t}\n\t\t\t}\n\n\t\t\tif (!response || !response.ok) {\n\t\t\t\tthrow lastError ?? new Error(\"Failed to get response after retries\");\n\t\t\t}\n\n\t\t\tlet started = false;\n\t\t\tconst ensureStarted = () => {\n\t\t\t\tif (!started) {\n\t\t\t\t\tstream.push({ type: \"start\", partial: output });\n\t\t\t\t\tstarted = true;\n\t\t\t\t}\n\t\t\t};\n\n\t\t\tconst resetOutput = () => {\n\t\t\t\toutput.content = [];\n\t\t\t\toutput.usage = {\n\t\t\t\t\tinput: 0,\n\t\t\t\t\toutput: 0,\n\t\t\t\t\tcacheRead: 0,\n\t\t\t\t\tcacheWrite: 0,\n\t\t\t\t\ttotalTokens: 0,\n\t\t\t\t\tcost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },\n\t\t\t\t};\n\t\t\t\toutput.stopReason = \"stop\";\n\t\t\t\toutput.errorMessage = undefined;\n\t\t\t\toutput.timestamp = Date.now();\n\t\t\t\tstarted = false;\n\t\t\t};\n\n\t\t\tconst streamResponse = async (activeResponse: Response): Promise<boolean> => {\n\t\t\t\tif (!activeResponse.body) {\n\t\t\t\t\tthrow new Error(\"No response body\");\n\t\t\t\t}\n\n\t\t\t\tlet hasContent = false;\n\t\t\t\tlet currentBlock: TextContent | ThinkingContent | null = null;\n\t\t\t\tconst blocks = output.content;\n\t\t\t\tconst blockIndex = () => blocks.length - 1;\n\n\t\t\t\t// Read SSE stream\n\t\t\t\tconst reader = activeResponse.body.getReader();\n\t\t\t\tconst decoder = new TextDecoder();\n\t\t\t\tlet buffer = \"\";\n\n\t\t\t\t// Set up abort handler to cancel reader when signal fires\n\t\t\t\tconst abortHandler = () => {\n\t\t\t\t\tvoid reader.cancel().catch(() => {});\n\t\t\t\t};\n\t\t\t\toptions?.signal?.addEventListener(\"abort\", abortHandler);\n\n\t\t\t\ttry {\n\t\t\t\t\twhile (true) {\n\t\t\t\t\t\t// Check abort signal before each read\n\t\t\t\t\t\tif (options?.signal?.aborted) {\n\t\t\t\t\t\t\tthrow new Error(\"Request was aborted\");\n\t\t\t\t\t\t}\n\n\t\t\t\t\t\tconst { done, value } = await reader.read();\n\t\t\t\t\t\tif (done) break;\n\n\t\t\t\t\t\tbuffer += decoder.decode(value, { stream: true });\n\t\t\t\t\t\tconst lines = buffer.split(\"\\n\");\n\t\t\t\t\t\tbuffer = lines.pop() || \"\";\n\n\t\t\t\t\t\tfor (const line of lines) {\n\t\t\t\t\t\t\tif (!line.startsWith(\"data:\")) continue;\n\n\t\t\t\t\t\t\tconst jsonStr = line.slice(5).trim();\n\t\t\t\t\t\t\tif (!jsonStr) continue;\n\n\t\t\t\t\t\t\tlet chunk: CloudCodeAssistResponseChunk;\n\t\t\t\t\t\t\ttry {\n\t\t\t\t\t\t\t\tchunk = JSON.parse(jsonStr);\n\t\t\t\t\t\t\t} catch {\n\t\t\t\t\t\t\t\tcontinue;\n\t\t\t\t\t\t\t}\n\n\t\t\t\t\t\t\t// Unwrap the response\n\t\t\t\t\t\t\tconst responseData = chunk.response;\n\t\t\t\t\t\t\tif (!responseData) continue;\n\t\t\t\t\t\t\t// Cloud Code Assist mirrors Gemini's responseId field. Keep the first non-empty one.\n\t\t\t\t\t\t\t// A single streamed response should retain the same ID across chunks.\n\t\t\t\t\t\t\toutput.responseId ||= responseData.responseId;\n\n\t\t\t\t\t\t\tconst candidate = responseData.candidates?.[0];\n\t\t\t\t\t\t\tif (candidate?.content?.parts) {\n\t\t\t\t\t\t\t\tfor (const part of candidate.content.parts) {\n\t\t\t\t\t\t\t\t\tif (part.text !== undefined) {\n\t\t\t\t\t\t\t\t\t\thasContent = true;\n\t\t\t\t\t\t\t\t\t\tconst isThinking = isThinkingPart(part);\n\t\t\t\t\t\t\t\t\t\tif (\n\t\t\t\t\t\t\t\t\t\t\t!currentBlock ||\n\t\t\t\t\t\t\t\t\t\t\t(isThinking && currentBlock.type !== \"thinking\") ||\n\t\t\t\t\t\t\t\t\t\t\t(!isThinking && currentBlock.type !== \"text\")\n\t\t\t\t\t\t\t\t\t\t) {\n\t\t\t\t\t\t\t\t\t\t\tif (currentBlock) {\n\t\t\t\t\t\t\t\t\t\t\t\tif (currentBlock.type === \"text\") {\n\t\t\t\t\t\t\t\t\t\t\t\t\tstream.push({\n\t\t\t\t\t\t\t\t\t\t\t\t\t\ttype: \"text_end\",\n\t\t\t\t\t\t\t\t\t\t\t\t\t\tcontentIndex: blocks.length - 1,\n\t\t\t\t\t\t\t\t\t\t\t\t\t\tcontent: currentBlock.text,\n\t\t\t\t\t\t\t\t\t\t\t\t\t\tpartial: output,\n\t\t\t\t\t\t\t\t\t\t\t\t\t});\n\t\t\t\t\t\t\t\t\t\t\t\t} else {\n\t\t\t\t\t\t\t\t\t\t\t\t\tstream.push({\n\t\t\t\t\t\t\t\t\t\t\t\t\t\ttype: \"thinking_end\",\n\t\t\t\t\t\t\t\t\t\t\t\t\t\tcontentIndex: blockIndex(),\n\t\t\t\t\t\t\t\t\t\t\t\t\t\tcontent: currentBlock.thinking,\n\t\t\t\t\t\t\t\t\t\t\t\t\t\tpartial: output,\n\t\t\t\t\t\t\t\t\t\t\t\t\t});\n\t\t\t\t\t\t\t\t\t\t\t\t}\n\t\t\t\t\t\t\t\t\t\t\t}\n\t\t\t\t\t\t\t\t\t\t\tif (isThinking) {\n\t\t\t\t\t\t\t\t\t\t\t\tcurrentBlock = { type: \"thinking\", thinking: \"\", thinkingSignature: undefined };\n\t\t\t\t\t\t\t\t\t\t\t\toutput.content.push(currentBlock);\n\t\t\t\t\t\t\t\t\t\t\t\tensureStarted();\n\t\t\t\t\t\t\t\t\t\t\t\tstream.push({\n\t\t\t\t\t\t\t\t\t\t\t\t\ttype: \"thinking_start\",\n\t\t\t\t\t\t\t\t\t\t\t\t\tcontentIndex: blockIndex(),\n\t\t\t\t\t\t\t\t\t\t\t\t\tpartial: output,\n\t\t\t\t\t\t\t\t\t\t\t\t});\n\t\t\t\t\t\t\t\t\t\t\t} else {\n\t\t\t\t\t\t\t\t\t\t\t\tcurrentBlock = { type: \"text\", text: \"\" };\n\t\t\t\t\t\t\t\t\t\t\t\toutput.content.push(currentBlock);\n\t\t\t\t\t\t\t\t\t\t\t\tensureStarted();\n\t\t\t\t\t\t\t\t\t\t\t\tstream.push({ type: \"text_start\", contentIndex: blockIndex(), partial: output });\n\t\t\t\t\t\t\t\t\t\t\t}\n\t\t\t\t\t\t\t\t\t\t}\n\t\t\t\t\t\t\t\t\t\tif (currentBlock.type === \"thinking\") {\n\t\t\t\t\t\t\t\t\t\t\tcurrentBlock.thinking += part.text;\n\t\t\t\t\t\t\t\t\t\t\tcurrentBlock.thinkingSignature = retainThoughtSignature(\n\t\t\t\t\t\t\t\t\t\t\t\tcurrentBlock.thinkingSignature,\n\t\t\t\t\t\t\t\t\t\t\t\tpart.thoughtSignature,\n\t\t\t\t\t\t\t\t\t\t\t);\n\t\t\t\t\t\t\t\t\t\t\tstream.push({\n\t\t\t\t\t\t\t\t\t\t\t\ttype: \"thinking_delta\",\n\t\t\t\t\t\t\t\t\t\t\t\tcontentIndex: blockIndex(),\n\t\t\t\t\t\t\t\t\t\t\t\tdelta: part.text,\n\t\t\t\t\t\t\t\t\t\t\t\tpartial: output,\n\t\t\t\t\t\t\t\t\t\t\t});\n\t\t\t\t\t\t\t\t\t\t} else {\n\t\t\t\t\t\t\t\t\t\t\tcurrentBlock.text += part.text;\n\t\t\t\t\t\t\t\t\t\t\tcurrentBlock.textSignature = retainThoughtSignature(\n\t\t\t\t\t\t\t\t\t\t\t\tcurrentBlock.textSignature,\n\t\t\t\t\t\t\t\t\t\t\t\tpart.thoughtSignature,\n\t\t\t\t\t\t\t\t\t\t\t);\n\t\t\t\t\t\t\t\t\t\t\tstream.push({\n\t\t\t\t\t\t\t\t\t\t\t\ttype: \"text_delta\",\n\t\t\t\t\t\t\t\t\t\t\t\tcontentIndex: blockIndex(),\n\t\t\t\t\t\t\t\t\t\t\t\tdelta: part.text,\n\t\t\t\t\t\t\t\t\t\t\t\tpartial: output,\n\t\t\t\t\t\t\t\t\t\t\t});\n\t\t\t\t\t\t\t\t\t\t}\n\t\t\t\t\t\t\t\t\t}\n\n\t\t\t\t\t\t\t\t\tif (part.functionCall) {\n\t\t\t\t\t\t\t\t\t\thasContent = true;\n\t\t\t\t\t\t\t\t\t\tif (currentBlock) {\n\t\t\t\t\t\t\t\t\t\t\tif (currentBlock.type === \"text\") {\n\t\t\t\t\t\t\t\t\t\t\t\tstream.push({\n\t\t\t\t\t\t\t\t\t\t\t\t\ttype: \"text_end\",\n\t\t\t\t\t\t\t\t\t\t\t\t\tcontentIndex: blockIndex(),\n\t\t\t\t\t\t\t\t\t\t\t\t\tcontent: currentBlock.text,\n\t\t\t\t\t\t\t\t\t\t\t\t\tpartial: output,\n\t\t\t\t\t\t\t\t\t\t\t\t});\n\t\t\t\t\t\t\t\t\t\t\t} else {\n\t\t\t\t\t\t\t\t\t\t\t\tstream.push({\n\t\t\t\t\t\t\t\t\t\t\t\t\ttype: \"thinking_end\",\n\t\t\t\t\t\t\t\t\t\t\t\t\tcontentIndex: blockIndex(),\n\t\t\t\t\t\t\t\t\t\t\t\t\tcontent: currentBlock.thinking,\n\t\t\t\t\t\t\t\t\t\t\t\t\tpartial: output,\n\t\t\t\t\t\t\t\t\t\t\t\t});\n\t\t\t\t\t\t\t\t\t\t\t}\n\t\t\t\t\t\t\t\t\t\t\tcurrentBlock = null;\n\t\t\t\t\t\t\t\t\t\t}\n\n\t\t\t\t\t\t\t\t\t\tconst providedId = part.functionCall.id;\n\t\t\t\t\t\t\t\t\t\tconst needsNewId =\n\t\t\t\t\t\t\t\t\t\t\t!providedId ||\n\t\t\t\t\t\t\t\t\t\t\toutput.content.some((b) => b.type === \"toolCall\" && b.id === providedId);\n\t\t\t\t\t\t\t\t\t\tconst toolCallId = needsNewId\n\t\t\t\t\t\t\t\t\t\t\t? `${part.functionCall.name}_${Date.now()}_${++toolCallCounter}`\n\t\t\t\t\t\t\t\t\t\t\t: providedId;\n\n\t\t\t\t\t\t\t\t\t\tconst toolCall: ToolCall = {\n\t\t\t\t\t\t\t\t\t\t\ttype: \"toolCall\",\n\t\t\t\t\t\t\t\t\t\t\tid: toolCallId,\n\t\t\t\t\t\t\t\t\t\t\tname: part.functionCall.name || \"\",\n\t\t\t\t\t\t\t\t\t\t\targuments: (part.functionCall.args as Record<string, unknown>) ?? {},\n\t\t\t\t\t\t\t\t\t\t\t...(part.thoughtSignature && { thoughtSignature: part.thoughtSignature }),\n\t\t\t\t\t\t\t\t\t\t};\n\n\t\t\t\t\t\t\t\t\t\toutput.content.push(toolCall);\n\t\t\t\t\t\t\t\t\t\tensureStarted();\n\t\t\t\t\t\t\t\t\t\tstream.push({ type: \"toolcall_start\", contentIndex: blockIndex(), partial: output });\n\t\t\t\t\t\t\t\t\t\tstream.push({\n\t\t\t\t\t\t\t\t\t\t\ttype: \"toolcall_delta\",\n\t\t\t\t\t\t\t\t\t\t\tcontentIndex: blockIndex(),\n\t\t\t\t\t\t\t\t\t\t\tdelta: JSON.stringify(toolCall.arguments),\n\t\t\t\t\t\t\t\t\t\t\tpartial: output,\n\t\t\t\t\t\t\t\t\t\t});\n\t\t\t\t\t\t\t\t\t\tstream.push({\n\t\t\t\t\t\t\t\t\t\t\ttype: \"toolcall_end\",\n\t\t\t\t\t\t\t\t\t\t\tcontentIndex: blockIndex(),\n\t\t\t\t\t\t\t\t\t\t\ttoolCall,\n\t\t\t\t\t\t\t\t\t\t\tpartial: output,\n\t\t\t\t\t\t\t\t\t\t});\n\t\t\t\t\t\t\t\t\t}\n\t\t\t\t\t\t\t\t}\n\t\t\t\t\t\t\t}\n\n\t\t\t\t\t\t\tif (candidate?.finishReason) {\n\t\t\t\t\t\t\t\toutput.stopReason = mapStopReasonString(candidate.finishReason);\n\t\t\t\t\t\t\t\tif (output.content.some((b) => b.type === \"toolCall\")) {\n\t\t\t\t\t\t\t\t\toutput.stopReason = \"toolUse\";\n\t\t\t\t\t\t\t\t}\n\t\t\t\t\t\t\t}\n\n\t\t\t\t\t\t\tif (responseData.usageMetadata) {\n\t\t\t\t\t\t\t\t// promptTokenCount includes cachedContentTokenCount, so subtract to get fresh input\n\t\t\t\t\t\t\t\tconst promptTokens = responseData.usageMetadata.promptTokenCount || 0;\n\t\t\t\t\t\t\t\tconst cacheReadTokens = responseData.usageMetadata.cachedContentTokenCount || 0;\n\t\t\t\t\t\t\t\toutput.usage = {\n\t\t\t\t\t\t\t\t\tinput: promptTokens - cacheReadTokens,\n\t\t\t\t\t\t\t\t\toutput:\n\t\t\t\t\t\t\t\t\t\t(responseData.usageMetadata.candidatesTokenCount || 0) +\n\t\t\t\t\t\t\t\t\t\t(responseData.usageMetadata.thoughtsTokenCount || 0),\n\t\t\t\t\t\t\t\t\tcacheRead: cacheReadTokens,\n\t\t\t\t\t\t\t\t\tcacheWrite: 0,\n\t\t\t\t\t\t\t\t\ttotalTokens: responseData.usageMetadata.totalTokenCount || 0,\n\t\t\t\t\t\t\t\t\tcost: {\n\t\t\t\t\t\t\t\t\t\tinput: 0,\n\t\t\t\t\t\t\t\t\t\toutput: 0,\n\t\t\t\t\t\t\t\t\t\tcacheRead: 0,\n\t\t\t\t\t\t\t\t\t\tcacheWrite: 0,\n\t\t\t\t\t\t\t\t\t\ttotal: 0,\n\t\t\t\t\t\t\t\t\t},\n\t\t\t\t\t\t\t\t};\n\t\t\t\t\t\t\t\tcalculateCost(model, output.usage);\n\t\t\t\t\t\t\t}\n\t\t\t\t\t\t}\n\t\t\t\t\t}\n\t\t\t\t} finally {\n\t\t\t\t\toptions?.signal?.removeEventListener(\"abort\", abortHandler);\n\t\t\t\t}\n\n\t\t\t\tif (currentBlock) {\n\t\t\t\t\tif (currentBlock.type === \"text\") {\n\t\t\t\t\t\tstream.push({\n\t\t\t\t\t\t\ttype: \"text_end\",\n\t\t\t\t\t\t\tcontentIndex: blockIndex(),\n\t\t\t\t\t\t\tcontent: currentBlock.text,\n\t\t\t\t\t\t\tpartial: output,\n\t\t\t\t\t\t});\n\t\t\t\t\t} else {\n\t\t\t\t\t\tstream.push({\n\t\t\t\t\t\t\ttype: \"thinking_end\",\n\t\t\t\t\t\t\tcontentIndex: blockIndex(),\n\t\t\t\t\t\t\tcontent: currentBlock.thinking,\n\t\t\t\t\t\t\tpartial: output,\n\t\t\t\t\t\t});\n\t\t\t\t\t}\n\t\t\t\t}\n\n\t\t\t\treturn hasContent;\n\t\t\t};\n\n\t\t\tlet receivedContent = false;\n\t\t\tlet currentResponse = response;\n\n\t\t\tfor (let emptyAttempt = 0; emptyAttempt <= MAX_EMPTY_STREAM_RETRIES; emptyAttempt++) {\n\t\t\t\tif (options?.signal?.aborted) {\n\t\t\t\t\tthrow new Error(\"Request was aborted\");\n\t\t\t\t}\n\n\t\t\t\tif (emptyAttempt > 0) {\n\t\t\t\t\tconst backoffMs = EMPTY_STREAM_BASE_DELAY_MS * 2 ** (emptyAttempt - 1);\n\t\t\t\t\tawait sleep(backoffMs, options?.signal);\n\n\t\t\t\t\tif (!requestUrl) {\n\t\t\t\t\t\tthrow new Error(\"Missing request URL\");\n\t\t\t\t\t}\n\n\t\t\t\t\tcurrentResponse = await fetch(requestUrl, {\n\t\t\t\t\t\tmethod: \"POST\",\n\t\t\t\t\t\theaders: requestHeaders,\n\t\t\t\t\t\tbody: requestBodyJson,\n\t\t\t\t\t\tsignal: options?.signal,\n\t\t\t\t\t});\n\n\t\t\t\t\tif (!currentResponse.ok) {\n\t\t\t\t\t\tconst retryErrorText = await currentResponse.text();\n\t\t\t\t\t\tthrow new Error(`Cloud Code Assist API error (${currentResponse.status}): ${retryErrorText}`);\n\t\t\t\t\t}\n\t\t\t\t}\n\n\t\t\t\tconst streamed = await streamResponse(currentResponse);\n\t\t\t\tif (streamed) {\n\t\t\t\t\treceivedContent = true;\n\t\t\t\t\tbreak;\n\t\t\t\t}\n\n\t\t\t\tif (emptyAttempt < MAX_EMPTY_STREAM_RETRIES) {\n\t\t\t\t\tresetOutput();\n\t\t\t\t}\n\t\t\t}\n\n\t\t\tif (!receivedContent) {\n\t\t\t\tthrow new Error(\"Cloud Code Assist API returned an empty response\");\n\t\t\t}\n\n\t\t\tif (options?.signal?.aborted) {\n\t\t\t\tthrow new Error(\"Request was aborted\");\n\t\t\t}\n\n\t\t\tif (output.stopReason === \"aborted\" || output.stopReason === \"error\") {\n\t\t\t\tthrow new Error(\"An unknown error occurred\");\n\t\t\t}\n\n\t\t\tstream.push({ type: \"done\", reason: output.stopReason, message: output });\n\t\t\tstream.end();\n\t\t} catch (error) {\n\t\t\tfor (const block of output.content) {\n\t\t\t\tif (\"index\" in block) {\n\t\t\t\t\tdelete (block as { index?: number }).index;\n\t\t\t\t}\n\t\t\t}\n\t\t\toutput.stopReason = options?.signal?.aborted ? \"aborted\" : \"error\";\n\t\t\toutput.errorMessage = error instanceof Error ? error.message : JSON.stringify(error);\n\t\t\tstream.push({ type: \"error\", reason: output.stopReason, error: output });\n\t\t\tstream.end();\n\t\t}\n\t})();\n\n\treturn stream;\n};\n\nexport const streamSimpleGoogleGeminiCli: StreamFunction<\"google-gemini-cli\", SimpleStreamOptions> = (\n\tmodel: Model<\"google-gemini-cli\">,\n\tcontext: Context,\n\toptions?: SimpleStreamOptions,\n): AssistantMessageEventStream => {\n\tconst apiKey = options?.apiKey;\n\tif (!apiKey) {\n\t\tthrow new Error(\"Google Cloud Code Assist requires OAuth authentication. Use /login to authenticate.\");\n\t}\n\n\tconst base = buildBaseOptions(model, options, apiKey);\n\tif (!options?.reasoning) {\n\t\treturn streamGoogleGeminiCli(model, context, {\n\t\t\t...base,\n\t\t\tthinking: { enabled: false },\n\t\t} satisfies GoogleGeminiCliOptions);\n\t}\n\n\tconst effort = clampReasoning(options.reasoning)!;\n\tif (isGemini3Model(model.id)) {\n\t\treturn streamGoogleGeminiCli(model, context, {\n\t\t\t...base,\n\t\t\tthinking: {\n\t\t\t\tenabled: true,\n\t\t\t\tlevel: getGeminiCliThinkingLevel(effort, model.id),\n\t\t\t},\n\t\t} satisfies GoogleGeminiCliOptions);\n\t}\n\n\tconst defaultBudgets: ThinkingBudgets = {\n\t\tminimal: 1024,\n\t\tlow: 2048,\n\t\tmedium: 8192,\n\t\thigh: 16384,\n\t};\n\tconst budgets = { ...defaultBudgets, ...options.thinkingBudgets };\n\n\tconst minOutputTokens = 1024;\n\tlet thinkingBudget = budgets[effort]!;\n\tconst maxTokens = Math.min((base.maxTokens || 0) + thinkingBudget, model.maxTokens);\n\n\tif (maxTokens <= thinkingBudget) {\n\t\tthinkingBudget = Math.max(0, maxTokens - minOutputTokens);\n\t}\n\n\treturn streamGoogleGeminiCli(model, context, {\n\t\t...base,\n\t\tmaxTokens,\n\t\tthinking: {\n\t\t\tenabled: true,\n\t\t\tbudgetTokens: thinkingBudget,\n\t\t},\n\t} satisfies GoogleGeminiCliOptions);\n};\n\nexport function buildRequest(\n\tmodel: Model<\"google-gemini-cli\">,\n\tcontext: Context,\n\tprojectId: string,\n\toptions: GoogleGeminiCliOptions = {},\n\tisAntigravity = false,\n): CloudCodeAssistRequest {\n\tconst contents = convertMessages(model, context);\n\n\tconst generationConfig: CloudCodeAssistRequest[\"request\"][\"generationConfig\"] = {};\n\tif (options.temperature !== undefined) {\n\t\tgenerationConfig.temperature = options.temperature;\n\t}\n\tif (options.maxTokens !== undefined) {\n\t\tgenerationConfig.maxOutputTokens = options.maxTokens;\n\t}\n\n\t// Thinking config\n\tif (options.thinking?.enabled && model.reasoning) {\n\t\tgenerationConfig.thinkingConfig = {\n\t\t\tincludeThoughts: true,\n\t\t};\n\t\t// Gemini 3 models use thinkingLevel, older models use thinkingBudget\n\t\tif (options.thinking.level !== undefined) {\n\t\t\t// Cast to any since our GoogleThinkingLevel mirrors Google's ThinkingLevel enum values\n\t\t\tgenerationConfig.thinkingConfig.thinkingLevel = options.thinking.level as any;\n\t\t} else if (options.thinking.budgetTokens !== undefined) {\n\t\t\tgenerationConfig.thinkingConfig.thinkingBudget = options.thinking.budgetTokens;\n\t\t}\n\t} else if (model.reasoning && options.thinking && !options.thinking.enabled) {\n\t\tgenerationConfig.thinkingConfig = getDisabledThinkingConfig(model.id);\n\t}\n\n\tconst request: CloudCodeAssistRequest[\"request\"] = {\n\t\tcontents,\n\t};\n\n\trequest.sessionId = options.sessionId;\n\n\t// System instruction must be object with parts, not plain string\n\tif (context.systemPrompt) {\n\t\trequest.systemInstruction = {\n\t\t\tparts: [{ text: sanitizeSurrogates(context.systemPrompt) }],\n\t\t};\n\t}\n\n\tif (Object.keys(generationConfig).length > 0) {\n\t\trequest.generationConfig = generationConfig;\n\t}\n\n\tif (context.tools && context.tools.length > 0) {\n\t\t// Claude models on Cloud Code Assist need the legacy `parameters` field;\n\t\t// the API translates it into Anthropic's `input_schema`.\n\t\tconst useParameters = model.id.startsWith(\"claude-\");\n\t\trequest.tools = convertTools(context.tools, useParameters);\n\t\tif (options.toolChoice) {\n\t\t\trequest.toolConfig = {\n\t\t\t\tfunctionCallingConfig: {\n\t\t\t\t\tmode: mapToolChoice(options.toolChoice),\n\t\t\t\t},\n\t\t\t};\n\t\t}\n\t}\n\n\tif (isAntigravity) {\n\t\tconst existingParts = request.systemInstruction?.parts ?? [];\n\t\trequest.systemInstruction = {\n\t\t\trole: \"user\",\n\t\t\tparts: [\n\t\t\t\t{ text: ANTIGRAVITY_SYSTEM_INSTRUCTION },\n\t\t\t\t{ text: `Please ignore following [ignore]${ANTIGRAVITY_SYSTEM_INSTRUCTION}[/ignore]` },\n\t\t\t\t...existingParts,\n\t\t\t],\n\t\t};\n\t}\n\n\treturn {\n\t\tproject: projectId,\n\t\tmodel: model.id,\n\t\trequest,\n\t\t...(isAntigravity ? { requestType: \"agent\" } : {}),\n\t\tuserAgent: isAntigravity ? \"antigravity\" : \"pi-coding-agent\",\n\t\trequestId: `${isAntigravity ? \"agent\" : \"pi\"}-${Date.now()}-${Math.random().toString(36).slice(2, 11)}`,\n\t};\n}\n\ntype ClampedThinkingLevel = Exclude<ThinkingLevel, \"xhigh\">;\n\nfunction getDisabledThinkingConfig(modelId: string): ThinkingConfig {\n\t// Google docs: Gemini 3.1 Pro cannot disable thinking, and Gemini 3 Flash / Flash-Lite\n\t// do not support full thinking-off either. For Gemini 3 models, use the lowest supported\n\t// thinkingLevel without includeThoughts so hidden thinking remains invisible to pi.\n\tif (isGemini3ProModel(modelId)) {\n\t\treturn { thinkingLevel: \"LOW\" as any };\n\t}\n\tif (isGemini3FlashModel(modelId)) {\n\t\treturn { thinkingLevel: \"MINIMAL\" as any };\n\t}\n\n\t// Gemini 2.x supports disabling via thinkingBudget = 0.\n\treturn { thinkingBudget: 0 };\n}\n\nfunction getGeminiCliThinkingLevel(effort: ClampedThinkingLevel, modelId: string): GoogleThinkingLevel {\n\tif (isGemini3ProModel(modelId)) {\n\t\tswitch (effort) {\n\t\t\tcase \"minimal\":\n\t\t\tcase \"low\":\n\t\t\t\treturn \"LOW\";\n\t\t\tcase \"medium\":\n\t\t\tcase \"high\":\n\t\t\t\treturn \"HIGH\";\n\t\t}\n\t}\n\tswitch (effort) {\n\t\tcase \"minimal\":\n\t\t\treturn \"MINIMAL\";\n\t\tcase \"low\":\n\t\t\treturn \"LOW\";\n\t\tcase \"medium\":\n\t\t\treturn \"MEDIUM\";\n\t\tcase \"high\":\n\t\t\treturn \"HIGH\";\n\t}\n}\n"]}